Disinformation* and glorification of violence in the context of the war in Ukraine
At Criteo, we strive to provide rigorous Supply Partner Guidelines to create a safe and efficient advertising solution that benefits our marketers and media owners. Our guidelines state that ”sharing or propagating false information or fabricated information designed to deceptively imitate news media content” is strictly prohibited.
Digital advertising is a key building block of the free and open internet. By generating most of the publishers’ revenue, digital advertising enables access to free information and, as such, supports independent journalism and access to quality information for internet users. Therefore, the indiscriminate limitation of ad displays on account of potential disinformation must be properly balanced against these benefits, or else we run the risk of unduly cutting publisher’s revenues and impinging on access to legitimate information and freedom of expression. Criteo’s efforts are thus a step forward toward a safer advertising experience for all of us, as well as a strong commitment against the monetization of disinformation.
We recognize that the spread of online disinformation is a genuine and complex problem: there is no 100% failproof solution that will be effective at the scale of the internet today. Just like in the wake of the Covid-19 pandemic, the recent tragic events in Ukraine were followed by a surge in conspiracy theories and articles spreading hate speech and disinformation about the conflict.
We empower our clients with granular controls to decide which inventory sources align with their brand. Our brand safety solution allows them to avoid topics that are not acceptable for their brand image, for example, “adult content”, “military” or “death”. Apart from using data from our external and internal disinformation reports, we continue to partner with Oracle Advertising and use their Potentially False Information (PFI) signals to identify inventory sources flagged as high risk of spreading disinformation. To obtain these new signals, Oracle collaborates with The Global Disinformation Index (GDI), “an independent not-for-profit organization dedicated to fighting disinformation on the Web and providing trusted, non-partisan and independent ratings to assess a site’s disinformation risk”.
Criteo, the leading Commerce Media Platform on the Open Internet, is committed to staying on the frontline of the fight against the monetization of disinformation; in this blog post, we explain how, thanks to the talent and resources of our AI Research Lab, our engineering teams have developed the first iteration of an Artificial Intelligence solution based on Natural Language Processing (NLP). It has already been instrumental for our Trust & Safety team in identifying and blocking inventory sources posing a high risk of spreading disinformation.
Website Content Analysis
At Criteo’s R&D, we extract contextual signals from the publishers’ content (websites). More precisely, we crawl half a billion publisher web pages daily (Criteo crawler) and analyze them using state-of-the-art machine learning (ML) techniques to extract relevant information.
A quick summary of how our main systems work: we crawl relevant web pages (by respecting the robots.txt specification), extract the text from them, and send it to several ML models. These models are based on NLP and were initially trained to leverage large multilingual datasets. In other words, these models understand human language and fall under the definition of “deep learning”. We further adapt them to our specific needs by a process called fine-tuning. For example, to classify web pages in IAB’s content taxonomy (Sports, News, etc.), we employ human annotators to manually label thousands of web pages. We then train a deep learning model for text classification on this data, and use the analysis as contextual features in other Criteo services, for example, recommendation, or contextual targeting.
In embeddings we trust
One of the byproducts of our NLP models is so-called page embeddings: we can extract from the model a numeric representation of each web page. More precisely, every page is represented by a vector of 100 numbers (in ML, we say ‘embedded into 100-dimensional space’, that is why we call them the ‘embeddings’). Why is this useful? If we have a representation of web pages expressed in numbers, we can compute the distance between two pages, which tells us how similar these pages are. For example, www.food.com is much closer to www.cookinglight.com than to www.eurosport.com/football.
Thanks to the embeddings, we can navigate the internet by following a given topic. Starting with one web page, for example, www.food.com/recipe/oven-crisp-chicken-wings, we can generate a list of web pages that are most similar to it. Why is this useful for tackling disinformation? We can focus on one specific disinformation topic and detect all the pages relevant to this topic. We only need to select a representative article that would be a good example of a web page relevant to the topic of conflict in Ukraine. Note that this step is crucial for detecting disinformation: due to the complexity of the task, models focusing on a specific topic have much better chances of performing well compared with general-purpose disinformation detectors. On the other hand, this might be one of the main limitations since every newly emerging topic requires a specific model.
The disinformation model
We have discussed how we can filter the internet on a particular topic thanks to the page embeddings. However, there are still too many pages about the conflict in Ukraine and it would be impossible for any corporate Trust & Safety team to manually review them all. We need to narrow down the list even more.
To select articles most likely to spread disinformation, we came up with a model that can estimate this probability from the web page embeddings. The brand safety team then manually reviewed articles with the highest probability of disinformation or propaganda.
The classification model is a three-layer neural network, a simpler architecture than the deep learning models mentioned above. Much of the prediction power stems from the deep learning embedding, used as input and carrying all the web page information.
To train the classifier (as a supervised classification task), we had to carefully build the training examples: we put pages corresponding to disinformation in one group and pages from trustworthy media in another group. The model then learns how to separate the two groups from these examples. We used websites already blocked by the Trust & Safety team as disinformation examples, and we manually listed trustworthy media such as cnn.com, lemonde.fr, theguardian.uk.
What were we able to achieve?
Thanks to this model, our Trust & Safety team can now proactively block content flagged for spreading disinformation. Our model makes the review process much more efficient by radically shortening the list of publishers for review: from hundreds of thousands to a few hundred publishers. In the submitted shortlist of more than 300 publishers, they all mentioned the topic of conflict in Ukraine. Trust & Safety found that one in two of these websites required further human review following our internal process. Thanks to our tool, there were 26 newly identified publishers that qualified for being blocked from Criteo’s inventory for infractions of our guidelines. These publishers were previously not identified by external providers. Interestingly, we have found not only Russian web pages but also US-based and some others since our model works across all languages.
As we have already mentioned, tagging exhaustively harmful websites is out of reach. Estimating reliable precision and recall for this task is challenging as well. Because the test data set was built manually, the excellent observed performances (> 0.95 AUROC) may not reflect those on the open internet. Nevertheless, the 26 blocked publishers were found to be ranked higher than average in the list of 300 reviewed domains, demonstrating the good performance of our model on previously unseen websites. We want to highlight the final incremental results on previously unseen websites, as the Trust & Safety team estimated.
Versatile tool applicable to any topic
This tool could be applied to the Ukraine conflict and any other critical global events to permanently block a publisher not complying with our Supply Partner Guidelines from our inventory. Crucially, our approach is designed to be easy to iterate on and adaptable to other topics since new topics prone to spreading disinformation may appear unpredictably. We only need to select the first web page representatives of the topic to build the topic classifier and provide examples of websites (with content related to the topic) from both trustworthy sources and sources spreading disinformation. The model setup and evaluation take only several hours, and the human review only a few days. This approach is applied to our whole inventory, making this a very quick solution to address these issues.
The building blocks of our engines in Criteo are ML-powered and many of them are of rich value if taken on their own. This project proved the power of the ML models we run at scale and the possibility of incorporating existing components into other projects.
We would like to acknowledge the cross-department work (R&D, Legal, Product Analytics) to make the Open Internet a safer and fairer place. This collaboration highlighted Criteo’s core values: “Open, Together, Impactful.”
If you are interested in more details about this work, you can check our research paper on this topic that we published in the ICML22 workshop Disinformation Countermeasures and Machine Learning (DisCoML). Check it out:
Multilingual Disinformation Detection for Digital Advertising
In today's world, the presence of online disinformation and propaganda is more widespread than ever.
arxiv.org
.* A disinformation website is a media that deceptively looks and feels like a trustworthy news outlet but lacks the editorial norms and processes for ensuring information accuracy, transparency, and credibility. It purposely spreads false or distorted information in a way that is sensational and emotionally charged for deceiving people for political, personal, or financial gain. ‘Disinformation’ is sometimes confused with ‘misinformation’, which is false information but is not deliberate.
Careers at Criteo | Criteo jobs
Find opportunities everywhere. Choose your next challenge.
careers.criteo.com
